You publish an invariant. Anvil proposes a patch, separate attackers try to break it, and Stone re-runs your real test harness with no model judgment. Accepting a patch becomes a contest against real tests. @the_Arow_H shared Temper.
What it does
You publish an invariant. Anvil proposes a patch, separate attackers try to break it, and Stone re-runs your real test harness with no model judgment. Accepting a patch becomes a contest against real tests. @the_Arow_H shared Temper. It runs when you ask it to, rather than on a timer.
Before you install
It needs access to your test harness; keep that scoped and review patches before merge. Open the share link and read the instructions it carries before you add it. @the_Arow_H shared this one publicly, and adding it copies the setup into your own account rather than giving you access to theirs.
Source
Public share: https://x.ai/bot/9Zp-JZWckPEjwBPp2EGtX. Announced by @the_Arow_H at https://x.com/the_Arow_H/status/2105392423417569434.
This bot was packaged and shared publicly by its author. Installing a shared bot means running someone else's instructions: review what it does, and revoke access if it oversteps. The install link is an official x.ai link — Grok Bot always shows you what will be added before confirming.